视觉感知的投影式三维语义分割后处理方法

doi:10.3969/j.issn.0255-8297.2026.04.008

摘要/Abstract

摘要： 投影式三维语义分割方法能够有效降低三维数据处理复杂度与计算开销，但现有方法大多依赖RGB颜色空间进行特征提取，易受光照变化、阴影干扰以及类别颜色相似性的影响，导致分割精度受限。针对上述问题，本文受人类视觉系统颜色感知机制启发，提出一种基于多颜色空间后处理的投影式三维语义分割方法。首先，通过二维投影与栅格化处理将三维场景转换为二维规则影像，并利用深度学习模型获得初始分割结果；其次，在低置信度区域引入更符合人类视觉感知特性的Lab与HSV颜色空间特征进行聚类优化，以提升类别可分性与区域一致性；再次，结合二维–三维映射关系，将优化后的语义标签恢复至三维场景，实现高精度语义标注。实验结果表明，本文方法在4组复杂城市场景数据上均取得较好的分割效果。相较传统聚类方法，总体分类精度（overall accuracy,OA）平均提升约21%,Kappa系数平均提升约0.28；相较单纯深度学习方法，OA平均提升约5%,Kappa系数平均提升约0.07。结果表明，该方法能够有效提升复杂场景下三维语义分割的精度与稳定性。

关键词: 三维语义分割, 投影式方法, 后处理优化, 深度学习

Abstract: Projection-based three-dimensional semantic segmentation methods can effectively reduce the processing complexity and computational cost of three-dimensional data. However, most existing methods rely on the RGB color space for feature extraction, making them susceptible to illumination variations, shadow interference, and category color similarity, which limits segmentation accuracy. To address these issues, inspired by the color perception mechanism of the human visual system, this paper proposed a projection-based three-dimensional semantic segmentation method based on multi-color space post-processing. First, the three-dimensional scene was transformed into a regular two-dimensional image through two-dimensional projection and rasterization, and an initial segmentation result was obtained using a deep learning model. Second, in low-confidence regions, Lab and HSV color space features that are more consistent with human visual perception characteristics were introduced for clustering optimization to improve category separability and regional consistency. Third, combined with the two-dimensional–threedimensional mapping relationship, the optimized semantic labels were restored to the threedimensional scene, achieving high-precision semantic annotation. The experimental results show that the proposed method achieves good segmentation performance on four sets of complex urban scene data. Compared with traditional clustering methods, the overall accuracy(OA) is improved by approximately 21% on average, and the Kappa coefficient is increased by approximately 0.28 on average. Compared with deep learning-only methods,the OA is improved by approximately 5% on average, and the Kappa coefficient is increased by approximately 0.07 on average. The results indicate that the proposed method can effectively enhance the accuracy and stability of three-dimensional semantic segmentation in complex scenes.

Key words: three-dimensional semantic segmentation, projection-based method, postprocessing optimization, deep learning

中图分类号:

P237

邹国镜, 丛铭, 崔建军, 韩玲. 视觉感知的投影式三维语义分割后处理方法[J]. 应用科学学报, 2026, 44(4): 626-643.

ZOU Guojing, CONG Ming, CUI Jianjun, HAN Ling. Visual Perception-Based Post-Processing Method for Projection-Based Three-Dimensional Semantic Segmentation[J]. Journal of Applied Sciences, 2026, 44(4): 626-643.

参考文献

[1] 郭震冬,吴昊,邵恒.大范围城市级实景三维模型建设及应用研究[J].江苏科技信息, 2025, 42(13):128-132.Guo Z D, Wu H, Shao H. Research on construction and application of large-scale urban real-scene3D model[J]. Jiangsu Science&Technology Information, 2025, 42(13):128-132.(in Chinese)
[2] 朱绪鹤,罗宁馨,王君毅.基于倾斜影像和机载LiDAR点云的城市级实景三维模型生产技术[J].测绘通报, 2024(2):144-147.Zhu X H, Luo N X, Wang J Y. Production technology of urban real-scene 3D model based on oblique images and airborne LiDAR point cloud[J]. Bulletin of Surveying and Mapping, 2024(2):144-147.(in Chinese)
[3] 李勇,佟国峰,杨景超,等.三维点云场景数据获取及其场景理解关键技术综述[J].激光与光电子学进展, 2019, 56(4):21-34.Li Y, Tong G F, Yang J C, et al. Review on key technologies of 3D point cloud scene data acquisition and scene understanding[J]. Laser&Optoelectronics Progress, 2019, 56(4):21-34.(in Chinese)
[4] Wang Y, Sun Y, Liu Z, et al. Dynamic graph CNN for learning on point clouds[J]. ACM Transactions on Graphics, 2019, 38(5):1-12.
[5] 仇志江,张林,姚垚,等.三维点云场景语义分割研究进展[J].中国图象图形学报, 2025, 30(7):2325-2342.Qiu Z J, Zhang L, Yao Y, et al. Research progress on semantic segmentation of 3D point cloud scenes[J]. Journal of Image and Graphics, 2025, 30(7):2325-2342.(in Chinese)
[6] Qi C R, Su H, Mo K, et al. Pointnet:deep learning on point sets for 3D classification and segmentation[C] //IEEE Conference on Computer Vision and Pattern Recognition, 2017:652-660.
[7] Qi C R, Yi L, Su H, et al. Pointnet++:deep hierarchical feature learning on point sets in a metric space[J]. Advances in Neural Information Processing Systems, 2017:30.
[8] Thomas H, Qi C R, Deschaud J E, et al. KPConv:flexible and deformable convolution for point clouds[C] //IEEE/CVF International Conference on Computer Vision. 2019:6411-6420.
[9] Hu Q, Yang B, Xie L, et al. Randla-net:efficient semantic segmentation of large-scale point clouds[C] //IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020:11108-11117.
[10] Hackel T, Savinov N, Ladicky L, et al. Semantic3d. net:a new large-scale point cloud classification benchmark[PP/OL]. V1(2017-04-12)[2026-01-30]. https://doi.org/10.48550/arXiv.1704.03847.
[11] Boulch A, Guerry J, Le Saux B, et al. SnapNet:3D point cloud semantic labeling with 2D deep segmentation networks[J]. Computers&Graphics, 2018, 71:189-198.
[12] Dai A, Nießner M. 3dmv:joint 3d-multi-view prediction for 3D semantic scene segmentation[C] //European Conference on Computer Vision(ECCV), 2018:452-468.
[13] Lawin F J, Danelljan M, Tosteberg P, et al. Deep projective 3D semantic segmentation[C] //International Conference on Computer Analysis of Images and Patterns, 2017:95-107.
[14] Rong M, Cui H, Shen S. Efficient 3D scene semantic segmentation via active learning on rendered2D images[J]. IEEE Transactions on Image Processing, 2023, 32:3521-3535.
[15] Dou J, Xue J, Fang J. SEG-VoxelNet for 3D vehicle detection from RGB and LiDAR data[C] //2019 International Conference on Robotics and Automation(ICRA). IEEE, 2019:4362-4368.
[16] Cheng B, Xu B, Deng Q, et al. MIFNet:multi-modal interactive fusion network for remote sensing semantic segmentation[C] //2025 IEEE International Geoscience and Remote Sensing Symposium, 2025:6980-6984.
[17] Achanta R, Shaji A, Smith K, et al. SLIC superpixels compared to state-of-the-art superpixel methods[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2012, 34(11):2274-2282.
[18] Ratajczak R, Crispim-Junior C, Fervers B, et al. Semantic segmentation refinement with deep edge superpixels to enhance historical land cover[C] //2020 IEEE International Geoscience and Remote Sensing Symposium, 2020:1813-1816.
[19] 许丽娜,肖奇,何鲁晓.考虑人类视觉特征的融合图像评价方法[J].武汉大学学报(信息科学版),2019, 44(4):546-554.Xu L N, Xiao Q, He L X. Fusion image evaluation method considering human visual characteristics[J]. Geomatics and Information Science of Wuhan University, 2019, 44(4):546-554.(in Chinese)
[20] Harrabi R, Braiek E B. Color image segmentation using a modified fuzzy C-means technique and different color spaces:application in the breast cancer cells images[C] //20141st International Conference on Advanced Technologies for Signal and Image Processing(ATSIP), 2014:231-236.
[21] Luo M R, Cui G, Rigg B. The development of the CIE 2000 colour-difference formula:CIEDE2000[J]. Color Research&Application, 2001, 26(5):340-350.
[22] 张蕾,李萄苑,李明飞,等.基于“面-廓-色”多通道特征融合的多肉叶片繁育评分算法[J].应用科学学报, 2026, 44(2):316-329.Zhang L, Li T Y, Li M F, et al. Succulent leaf breeding scoring algorithm based on“surfacecontour-color” multi-channel feature fusion[J]. Journal of Applied Sciences, 2026, 44(2):316-329.(in Chinese)